Bundle Method for optimizing convex and non convex risks.¶
This code implements an algorithm named NRBM (Non convex Regularized Bundle Methods) that we designed to deal efficiently with regularized non convex risks as often encountered in the machine learning field. It is a generic code able to deal with convex, non convex, smooth and non smooth risks.The code is available for download:
Please cite either the ICML or the JMLR paper below if you use our code (ICML bibtex).
Three documents are related to the method.
- A ICML 2009 (Intenational Conference on Machine Learning) paper
The ICML paper includes a brief description of the algorithm and of results for learning Hidden Markov Models with a large margin criterion.
- A JMLR 2012 (Journal of Machine Learning Research) paper
The JMLR article provides more details on the algorithm and report empirical evaluation on many real machine learning problems.
- An internal report at LIP6/UPMC
The internal report provides additional theoretical results related to the convergence analysis of algorithm NRBM.
- Site web: https://forge.lip6.fr/projects/nrbm
Manager : Thierry Artieres